Elo difference for 4-8-16-24-32 cores

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

x3
Posts: 13
Joined: Mon Sep 08, 2014 6:54 pm

Elo difference for 4-8-16-24-32 cores

Post by x3 »

SF5 8-16 core 37 elo 3200 games Xeon
SF5 16-24 core 7 elo 1600 games. Opteron
3m+1, and m4+1 8 move book repeating.

Anyone with other programs i.e. 4 to 8 cores?!

x3
bob
Posts: 20943
Joined: Mon Feb 27, 2006 7:30 pm
Location: Birmingham, AL

Re: Elo difference for 4-8-16-24-32 cores

Post by bob »

x3 wrote:SF5 8-16 core 37 elo 3200 games Xeon
SF5 16-24 core 7 elo 1600 games. Opteron
3m+1, and m4+1 8 move book repeating.

Anyone with other programs i.e. 4 to 8 cores?!

x3
Are these self-test are against other opponents? If self-test, the Elos are going to be significantly inflated. If against other opponents, 1600 games is not enough to be confident of +7 Elo at all.
x3
Posts: 13
Joined: Mon Sep 08, 2014 6:54 pm

Re: Elo difference for 4-8-16-24-32 cores

Post by x3 »

No it is the same Engine, but prove if it is significantly different against other Engines. 2 1600 games is not enough but it indicates that the difference is rather small.

x3
Sedat Canbaz
Posts: 3018
Joined: Thu Mar 09, 2006 11:58 am
Location: Antalya/Turkey

Re: Elo difference for 4-8-16-24-32 cores

Post by Sedat Canbaz »

x3 wrote:SF5 8-16 core 37 elo 3200 games Xeon
SF5 16-24 core 7 elo 1600 games. Opteron
3m+1, and m4+1 8 move book repeating.

Anyone with other programs i.e. 4 to 8 cores?!

x3

Very useful hardware testing... thanks !

It seems many MP engines (including Stockfish) suffer in case of using 16 cores or 24 cores!
I mean they don't gain much Elo...
And probably it will take some time to see the top mp engines to play at maximum performance
This is one of the main reasons about why I still use Intel i7 six-core machines

And during my hardware testings I noticed that,
In case of doubling the processor speed usually the mp engines perform around 50-90 Elo points
Of course, I mean for 1 core, 2 cores vs 4 cores, 6 cores, 8 cores

BTW, can we download the games?
I wonder about the kns values and conditions too
-I think the games are played on same PC, right ? (I mean it's not Auto232 mode)
-Under which GUI ?!
-Which book? and it was learning off or on (if under Fritz, Shredder Classic GUIs) ?
-Hashtable size ?
-Ponder OFF ? (I think)
-Endgames ?
etc...


Best,
Sedat
Sedat Canbaz
Posts: 3018
Joined: Thu Mar 09, 2006 11:58 am
Location: Antalya/Turkey

Re: Elo difference for 4-8-16-24-32 cores

Post by Sedat Canbaz »

x3 wrote: Anyone with other programs i.e. 4 to 8 cores?!
x3
Image

And here is another hardware speed Elo test !

Note: Ponder ON matches; engine matches between 2 PCs, mainly in Auto232 mode, some of the games are played on Playchess)

All engines are played with same neutral short book (one of my Perfect book series)

The time control was Blitz (3 minutes per side)

Not with Stockfish, but Rybka 8 core performed very very good...
Of course Rybka 8 core played 200 games are its not enough data....
But at least the the published table shows us that Rybka 3 on Majd's 8 core machine was very powerful ;)

Best,
Sedat
x3
Posts: 13
Joined: Mon Sep 08, 2014 6:54 pm

Re: Elo difference for 4-8-16-24-32 cores

Post by x3 »

Xeon 3,4 GHz
Ponder off
My book 8 moves repeating, no Learning.
Shredder the best GUI.
Hash I do not remember 1 or 2 gb.

Btw I am now running an equal test with Komodo, and so far it performs slightly better in both categories, only 400 games so far, about +10 vs SF5.

x3.
mjlef
Posts: 1494
Joined: Thu Mar 30, 2006 2:08 pm

Re: Elo difference for 4-8-16-24-32 cores

Post by mjlef »

x3 wrote:Xeon 3,4 GHz
Ponder off
My book 8 moves repeating, no Learning.
Shredder the best GUI.
Hash I do not remember 1 or 2 gb.

Btw I am now running an equal test with Komodo, and so far it performs slightly better in both categories, only 400 games so far, about +10 vs SF5.

x3.
We really appreciate you running Komodo 8, and we will be fascinated to see how it scales elo-wise with more threads/cores.

Mark
User avatar
Leto
Posts: 2071
Joined: Thu May 04, 2006 3:40 am
Location: Dune

Re: Elo difference for 4-8-16-24-32 cores

Post by Leto »

How does Houdini 4 perform with more than 4 cores? In my CEGT 40/4 testing it scales horribly above 4 cores, no performance increase.
User avatar
Leto
Posts: 2071
Joined: Thu May 04, 2006 3:40 am
Location: Dune

Re: Elo difference for 4-8-16-24-32 cores

Post by Leto »

mjlef wrote:
x3 wrote:Xeon 3,4 GHz
Ponder off
My book 8 moves repeating, no Learning.
Shredder the best GUI.
Hash I do not remember 1 or 2 gb.

Btw I am now running an equal test with Komodo, and so far it performs slightly better in both categories, only 400 games so far, about +10 vs SF5.

x3.
We really appreciate you running Komodo 8, and we will be fascinated to see how it scales elo-wise with more threads/cores.

Mark
From my testing at CEGT 40/4 Komodo 8 scales very well up to 8 cores, and then I noticed no performance gain above 8 cores. More games are needed and I have some gauntlets with Komodo 8, Stockfish 5, Gull 3, and Rybka 4 completed that will be added to the list soon.

Here's what we have so far from the list (does not included the new gauntlets for 12 cores and 8 cores I've completed since). Also I will start working on 4 cores today on my three computers:

Komodo 8 1 core: 3151
Komodo 8 8 cores: 3298
Komodo 8 12 cores: 3291
Gain from 1 core to 8 cores: 147 elo
Gain from 8 core to 12 cores: -7 elo
Gain from 1 core to 12 cores: 140 elo

In comparison here's Stockfish 5:
Stockfish 5 1 core: 3157
Stockfish 5 4 cores: 3228
Stockfish 5 8 cores: 3304
Stockfish 5 12 cores: 3313
Gain from 1 core to 4 cores: 71 elo
Gain from 4 core to 8 cores: 76 elo
Gain from 8 cores to 12 cores: 9 elo
Gain from 1 core to 8 cores: 147 elo
Gain from 1 cores to 12 cores: 156 elo

In another comparison here's Rybka 4:
Rybka 4 1 core: 2984
Rybka 4 4 cores: 3071
Rybka 4 8 cores: 3121
Rybka 4 12 cores: 3105 (based only on 50 games, more will be added soon)
Gain from 1 core to 4 cores: 87 elo
Gain from 4 cores to 8 cores: 50 elo
Gain from 8 cores to 12 cores: -16 elo (based only on 50 games)
Gain from 1 core to 8 cores: 137 elo
Gain from 1 core to 12 cores: 121 elo

And Houdini 4:
Houdini 4 1 core: 3133
Houdini 4 4 cores: 3220
Houdini 4 8 cores: 3225
Gain from 1 core to 4 cores: 87 elo
Gain from 4 cores to 8 cores: 5 elo
Gain from 1 core to 8 cores: 92 elo
Modern Times
Posts: 3546
Joined: Thu Jun 07, 2012 11:02 pm

Re: Elo difference for 4-8-16-24-32 cores

Post by Modern Times »

I could be wrong, but I don't think blitz is likely to bring out the full advantage of lots of cores. You're more likely to see it with a longer time control. But as always, we need to make compromises as we don't have unlimited resources.